Methodology of Quantitative Land Cover Characterization from MODIS, a case study in Northeast China.

WANG Zhengxing
Institute of Geographical Sciences and Natural Resources Research (IGSNRR)
Chinese Academy of Sciences (CAS)

Repeatability is one of the key principles in classic science. It means a fact or theory holds, only if anyone can repeat it, given certain data and processing guidelines. It is not sufficient for the first discoverer to prove it himself. This has been regarded as a touchstone in scientific community for a few hundred years. Yet value of repeatability has not been fully appreciated in remote sensing application: given same raw data, given same land cover system, different people may produce various land cover maps. This is a serious problem for long-term global change monitoring because only real change is interested, any other changes from investigators should be avoided as much as possible. Three factors are required to repeatability: data consistency, transparent and objective processing, and quantitative characterization of land cover types. To investigate the degree of repeatability, the full year 16-day MODIS-NDVI and NDWI time serials in 2002 are analyzed based on more than 60 ground samples. The major efforts include follows. 

1. Remote Sensing Data Consistency and Re-Building. Data from various sources are  "consistent" if they have a stable, quantitative relation. Consistent data should have detailed quality assessment and all noises should be removed, or flagged. With few exceptions, the temporal and spatial scale of global change monitoring makes the use of problematic data a common practice. For instance, the MODIS vegetation indices, much better than AVHRR-NDVI after 20 year improvements, does not mean too good to need any pre-processing. In fact, obviously there are some noises and these inconsistency should be detected and removed in a objective method. Based on the time serious and relation between different parameters, a supervised approach is used to re-build 16-day MODIS time series in 2002. The output include NDVI, red, NIR, SWIR, blue, and Normalized Difference Water Index (NDWI). Only NDVI and NDWI were used for further investigation.

2. Quantitative Land Cover Characterization from MODIS-NDVI Time Serials. NDVI time serials are used to quantitatively characterize 15 land covers (LCs) in Northeastern China. The relation between NDVI and LCs falls into three categories:

In the first group, sole NDVI time serials are sufficient to characterize LCs. A good example is forest, which means yearly average NDVI>0.45, or average NDVI>0.65 during growing season (day113-273), or average NDVI>0.55 during green-up (day113-145). Forest could be further characterized into three sub LCs using average NDVI during winter (Day1-97, Day289-353): NDVI>=0.53 for evergreen, NDVI=0.53~0.38 for mixed forest, and NDVI<=0.38 for deciduous forest

In the second category, only NDVI time serials are insufficient to characterize LCs, but is necessary for next-step investigation. For instance, NDVI may not differentiate some irrigated arable land and rain-fed arable land, since both of them may have similar seasonality and NDVI values. But NDVI could easily split arable land from natural vegetation, the latter has a 30-day earlier green-up in NDVI temporal profile.    

The third group include city and water. NDVI time serials is not a good parameter for water since the composite algorithm is "Vegetation Oriented". Clouds have more chance to be remained over water bodies because they have larger NDVI than water. In the case of city, it share common NDVI character with bare soil and sparse vegetation, and have many possibilities for different cities. Therefore, the best policy is to borrow these information from more reliable sources.    

3. Quantitative Land Cover Characterization from MODIS-NDWI Time Serials.  Normally, NDWI has close relation with NDVI in vegetated region during growing season, yet NDWI could supply independent information when vegetation is sparse or the background is rich in water. The NDVI similarity problem for irrigated and rain-fed is very simple for NDWI, because irrigated arable land has very wet ground in early period when crops are still very small, which help to segment two LCs by NDWI. It should be mentioned here that NDWI may have some potentials in broader fields, such us monitoring vegetation heath, or early warning of wild fires.    

4. Quantitatively Characterize Land Covers With a Decision Tree. Finally, a decision tree is constructed based on rules obtained in previous analyses. 13 land cover types were classified based on quantitative parameters. Full validation of the classification is not yet finished, but for LCs with reliable ground information, the results are promising. 

5. Conclusion.  To conclude, for consistent MODIS-NDVI and NDWI, there is a reliably quantitative relation between some LCs and remote sensing parameters. The degree of repeatability depends both on the quality of remote sensing data and the understanding of ground condition. A better knowledge of both will facilitate more reliable, more quantitative investigation of land cover changes.

Key words: MODIS, NDVI, NDWI, Land Cover Classification, Repeatability